fix(gguf): fall back to expert_feed_forward_length for MoE-only configs

[mvkorobkov]

Summary

DeepSeek-V4 family (and other MoE-only GGUFs) emit only `{arch}.expert_feed_forward_length` and never set the global `{arch}.feed_forward_length` — there's no dense FFN layer above the per-expert size. The current loader reads only the global key, so `intermediate_size` comes back as 0 and config validation rejects the model:

```
$ larql extract --level browse \
DeepSeek-V4-Flash-Q3_K_M-00001-of-00003.gguf
Error: failed to load GGUF model: config validation failed:
[ConfigValidationError { field: "intermediate_size",
message: "must be greater than 0" }]
```

Fix

In every llama.cpp-supported MoE architecture the per-expert FFN inner dim matches what a non-MoE variant would call `intermediate_size` (DS-V2-Lite-Chat happens to emit both at the same value when MoE is active). Fall back to the per-expert key only when the global key is absent.

```rust
let intermediate_size = {
let global = get_arch_u32(GGUF_FEED_FORWARD_LENGTH);
if global > 0 { global } else { get_arch_u32(GGUF_EXPERT_FEED_FORWARD_LENGTH) }
};
```

Tests

• `test_gguf_to_config_json_falls_back_to_expert_feed_forward_length_on_moe`: synthesises DS-V4-Flash-shaped metadata (no `feed_forward_length`, `expert_feed_forward_length: 2048`); verifies `intermediate_size: 2048` and that the validated detection path now accepts the config.
• `test_gguf_to_config_json_prefers_global_feed_forward_length_when_both_present`: when both keys are emitted, global still wins — no behaviour change for non-MoE-only models.

281/281 larql-models tests pass.

Context

Part of the DeepSeek-V2/V3/Kimi-K2/V4 GGUF chain I've been working through: #133 (merged) for GGUF input, #135 for MLA metadata, #136 for multi-shard splits, #137 for the MLA capabilities gate. This is the last "small" piece — the only remaining blocker after this is the per-tensor streaming GGUF reader needed to fit 500 GB+ models in available RAM, which is a separate larger PR.